xref: /linux/drivers/infiniband/hw/hns/hns_roce_qp.c (revision 4f58e6dceb0e44ca8f21568ed81e1df24e55964c)
1 /*
2  * Copyright (c) 2016 Hisilicon Limited.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/platform_device.h>
35 #include <rdma/ib_umem.h>
36 #include "hns_roce_common.h"
37 #include "hns_roce_device.h"
38 #include "hns_roce_hem.h"
39 #include "hns_roce_user.h"
40 
41 #define DB_REG_OFFSET			0x1000
42 #define SQP_NUM				12
43 
44 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
45 {
46 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
47 	struct device *dev = &hr_dev->pdev->dev;
48 	struct hns_roce_qp *qp;
49 
50 	spin_lock(&qp_table->lock);
51 
52 	qp = __hns_roce_qp_lookup(hr_dev, qpn);
53 	if (qp)
54 		atomic_inc(&qp->refcount);
55 
56 	spin_unlock(&qp_table->lock);
57 
58 	if (!qp) {
59 		dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
60 		return;
61 	}
62 
63 	qp->event(qp, (enum hns_roce_event)event_type);
64 
65 	if (atomic_dec_and_test(&qp->refcount))
66 		complete(&qp->free);
67 }
68 
69 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
70 				 enum hns_roce_event type)
71 {
72 	struct ib_event event;
73 	struct ib_qp *ibqp = &hr_qp->ibqp;
74 
75 	if (ibqp->event_handler) {
76 		event.device = ibqp->device;
77 		event.element.qp = ibqp;
78 		switch (type) {
79 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
80 			event.event = IB_EVENT_PATH_MIG;
81 			break;
82 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
83 			event.event = IB_EVENT_COMM_EST;
84 			break;
85 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
86 			event.event = IB_EVENT_SQ_DRAINED;
87 			break;
88 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
89 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
90 			break;
91 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
92 			event.event = IB_EVENT_QP_FATAL;
93 			break;
94 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
95 			event.event = IB_EVENT_PATH_MIG_ERR;
96 			break;
97 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
98 			event.event = IB_EVENT_QP_REQ_ERR;
99 			break;
100 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
101 			event.event = IB_EVENT_QP_ACCESS_ERR;
102 			break;
103 		default:
104 			dev_dbg(ibqp->device->dma_device, "roce_ib: Unexpected event type %d on QP %06lx\n",
105 				type, hr_qp->qpn);
106 			return;
107 		}
108 		ibqp->event_handler(&event, ibqp->qp_context);
109 	}
110 }
111 
112 static int hns_roce_reserve_range_qp(struct hns_roce_dev *hr_dev, int cnt,
113 				     int align, unsigned long *base)
114 {
115 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
116 	int ret = 0;
117 	unsigned long qpn;
118 
119 	ret = hns_roce_bitmap_alloc_range(&qp_table->bitmap, cnt, align, &qpn);
120 	if (ret == -1)
121 		return -ENOMEM;
122 
123 	*base = qpn;
124 
125 	return 0;
126 }
127 
128 enum hns_roce_qp_state to_hns_roce_state(enum ib_qp_state state)
129 {
130 	switch (state) {
131 	case IB_QPS_RESET:
132 		return HNS_ROCE_QP_STATE_RST;
133 	case IB_QPS_INIT:
134 		return HNS_ROCE_QP_STATE_INIT;
135 	case IB_QPS_RTR:
136 		return HNS_ROCE_QP_STATE_RTR;
137 	case IB_QPS_RTS:
138 		return HNS_ROCE_QP_STATE_RTS;
139 	case IB_QPS_SQD:
140 		return HNS_ROCE_QP_STATE_SQD;
141 	case IB_QPS_ERR:
142 		return HNS_ROCE_QP_STATE_ERR;
143 	default:
144 		return HNS_ROCE_QP_NUM_STATE;
145 	}
146 }
147 
148 static int hns_roce_gsi_qp_alloc(struct hns_roce_dev *hr_dev, unsigned long qpn,
149 				 struct hns_roce_qp *hr_qp)
150 {
151 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
152 	int ret;
153 
154 	if (!qpn)
155 		return -EINVAL;
156 
157 	hr_qp->qpn = qpn;
158 
159 	spin_lock_irq(&qp_table->lock);
160 	ret = radix_tree_insert(&hr_dev->qp_table_tree,
161 				hr_qp->qpn & (hr_dev->caps.num_qps - 1), hr_qp);
162 	spin_unlock_irq(&qp_table->lock);
163 	if (ret) {
164 		dev_err(&hr_dev->pdev->dev, "QPC radix_tree_insert failed\n");
165 		goto err_put_irrl;
166 	}
167 
168 	atomic_set(&hr_qp->refcount, 1);
169 	init_completion(&hr_qp->free);
170 
171 	return 0;
172 
173 err_put_irrl:
174 
175 	return ret;
176 }
177 
178 static int hns_roce_qp_alloc(struct hns_roce_dev *hr_dev, unsigned long qpn,
179 			     struct hns_roce_qp *hr_qp)
180 {
181 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
182 	struct device *dev = &hr_dev->pdev->dev;
183 	int ret;
184 
185 	if (!qpn)
186 		return -EINVAL;
187 
188 	hr_qp->qpn = qpn;
189 
190 	/* Alloc memory for QPC */
191 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
192 	if (ret) {
193 		dev_err(dev, "QPC table get failed\n");
194 		goto err_out;
195 	}
196 
197 	/* Alloc memory for IRRL */
198 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
199 	if (ret) {
200 		dev_err(dev, "IRRL table get failed\n");
201 		goto err_put_qp;
202 	}
203 
204 	spin_lock_irq(&qp_table->lock);
205 	ret = radix_tree_insert(&hr_dev->qp_table_tree,
206 				hr_qp->qpn & (hr_dev->caps.num_qps - 1), hr_qp);
207 	spin_unlock_irq(&qp_table->lock);
208 	if (ret) {
209 		dev_err(dev, "QPC radix_tree_insert failed\n");
210 		goto err_put_irrl;
211 	}
212 
213 	atomic_set(&hr_qp->refcount, 1);
214 	init_completion(&hr_qp->free);
215 
216 	return 0;
217 
218 err_put_irrl:
219 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
220 
221 err_put_qp:
222 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
223 
224 err_out:
225 	return ret;
226 }
227 
228 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
229 {
230 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
231 	unsigned long flags;
232 
233 	spin_lock_irqsave(&qp_table->lock, flags);
234 	radix_tree_delete(&hr_dev->qp_table_tree,
235 			  hr_qp->qpn & (hr_dev->caps.num_qps - 1));
236 	spin_unlock_irqrestore(&qp_table->lock, flags);
237 }
238 
239 void hns_roce_qp_free(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
240 {
241 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
242 
243 	if (atomic_dec_and_test(&hr_qp->refcount))
244 		complete(&hr_qp->free);
245 	wait_for_completion(&hr_qp->free);
246 
247 	if ((hr_qp->ibqp.qp_type) != IB_QPT_GSI) {
248 		hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
249 		hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
250 	}
251 }
252 
253 void hns_roce_release_range_qp(struct hns_roce_dev *hr_dev, int base_qpn,
254 			       int cnt)
255 {
256 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
257 
258 	if (base_qpn < (hr_dev->caps.sqp_start + 2 * hr_dev->caps.num_ports))
259 		return;
260 
261 	hns_roce_bitmap_free_range(&qp_table->bitmap, base_qpn, cnt);
262 }
263 
264 static int hns_roce_set_rq_size(struct hns_roce_dev *hr_dev,
265 				struct ib_qp_cap *cap, int is_user, int has_srq,
266 				struct hns_roce_qp *hr_qp)
267 {
268 	u32 max_cnt;
269 	struct device *dev = &hr_dev->pdev->dev;
270 
271 	/* Check the validity of QP support capacity */
272 	if (cap->max_recv_wr > hr_dev->caps.max_wqes ||
273 	    cap->max_recv_sge > hr_dev->caps.max_rq_sg) {
274 		dev_err(dev, "RQ WR or sge error!max_recv_wr=%d max_recv_sge=%d\n",
275 			cap->max_recv_wr, cap->max_recv_sge);
276 		return -EINVAL;
277 	}
278 
279 	/* If srq exit, set zero for relative number of rq */
280 	if (has_srq) {
281 		if (cap->max_recv_wr) {
282 			dev_dbg(dev, "srq no need config max_recv_wr\n");
283 			return -EINVAL;
284 		}
285 
286 		hr_qp->rq.wqe_cnt = hr_qp->rq.max_gs = 0;
287 	} else {
288 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge)) {
289 			dev_err(dev, "user space no need config max_recv_wr max_recv_sge\n");
290 			return -EINVAL;
291 		}
292 
293 		/* In v1 engine, parameter verification procession */
294 		max_cnt = cap->max_recv_wr > HNS_ROCE_MIN_WQE_NUM ?
295 			  cap->max_recv_wr : HNS_ROCE_MIN_WQE_NUM;
296 		hr_qp->rq.wqe_cnt = roundup_pow_of_two(max_cnt);
297 
298 		if ((u32)hr_qp->rq.wqe_cnt > hr_dev->caps.max_wqes) {
299 			dev_err(dev, "hns_roce_set_rq_size rq.wqe_cnt too large\n");
300 			return -EINVAL;
301 		}
302 
303 		max_cnt = max(1U, cap->max_recv_sge);
304 		hr_qp->rq.max_gs = roundup_pow_of_two(max_cnt);
305 		/* WQE is fixed for 64B */
306 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz);
307 	}
308 
309 	cap->max_recv_wr = hr_qp->rq.max_post = hr_qp->rq.wqe_cnt;
310 	cap->max_recv_sge = hr_qp->rq.max_gs;
311 
312 	return 0;
313 }
314 
315 static int hns_roce_set_user_sq_size(struct hns_roce_dev *hr_dev,
316 				     struct hns_roce_qp *hr_qp,
317 				     struct hns_roce_ib_create_qp *ucmd)
318 {
319 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
320 	u8 max_sq_stride = ilog2(roundup_sq_stride);
321 
322 	/* Sanity check SQ size before proceeding */
323 	if ((u32)(1 << ucmd->log_sq_bb_count) > hr_dev->caps.max_wqes ||
324 	     ucmd->log_sq_stride > max_sq_stride ||
325 	     ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
326 		dev_err(&hr_dev->pdev->dev, "check SQ size error!\n");
327 		return -EINVAL;
328 	}
329 
330 	hr_qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count;
331 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
332 
333 	/* Get buf size, SQ and RQ  are aligned to page_szie */
334 	hr_qp->buff_size = HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt <<
335 					     hr_qp->rq.wqe_shift), PAGE_SIZE) +
336 			   HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
337 					     hr_qp->sq.wqe_shift), PAGE_SIZE);
338 
339 	hr_qp->sq.offset = 0;
340 	hr_qp->rq.offset = HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
341 					     hr_qp->sq.wqe_shift), PAGE_SIZE);
342 
343 	return 0;
344 }
345 
346 static int hns_roce_set_kernel_sq_size(struct hns_roce_dev *hr_dev,
347 				       struct ib_qp_cap *cap,
348 				       enum ib_qp_type type,
349 				       struct hns_roce_qp *hr_qp)
350 {
351 	struct device *dev = &hr_dev->pdev->dev;
352 	u32 max_cnt;
353 	(void)type;
354 
355 	if (cap->max_send_wr  > hr_dev->caps.max_wqes  ||
356 	    cap->max_send_sge > hr_dev->caps.max_sq_sg ||
357 	    cap->max_inline_data > hr_dev->caps.max_sq_inline) {
358 		dev_err(dev, "hns_roce_set_kernel_sq_size error1\n");
359 		return -EINVAL;
360 	}
361 
362 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
363 	hr_qp->sq_max_wqes_per_wr = 1;
364 	hr_qp->sq_spare_wqes = 0;
365 
366 	/* In v1 engine, parameter verification procession */
367 	max_cnt = cap->max_send_wr > HNS_ROCE_MIN_WQE_NUM ?
368 		  cap->max_send_wr : HNS_ROCE_MIN_WQE_NUM;
369 	hr_qp->sq.wqe_cnt = roundup_pow_of_two(max_cnt);
370 	if ((u32)hr_qp->sq.wqe_cnt > hr_dev->caps.max_wqes) {
371 		dev_err(dev, "hns_roce_set_kernel_sq_size sq.wqe_cnt too large\n");
372 		return -EINVAL;
373 	}
374 
375 	/* Get data_seg numbers */
376 	max_cnt = max(1U, cap->max_send_sge);
377 	hr_qp->sq.max_gs = roundup_pow_of_two(max_cnt);
378 
379 	/* Get buf size, SQ and RQ  are aligned to page_szie */
380 	hr_qp->buff_size = HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt <<
381 					     hr_qp->rq.wqe_shift), PAGE_SIZE) +
382 			   HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
383 					     hr_qp->sq.wqe_shift), PAGE_SIZE);
384 	hr_qp->sq.offset = 0;
385 	hr_qp->rq.offset = HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
386 					      hr_qp->sq.wqe_shift), PAGE_SIZE);
387 
388 	/* Get wr and sge number which send */
389 	cap->max_send_wr = hr_qp->sq.max_post = hr_qp->sq.wqe_cnt;
390 	cap->max_send_sge = hr_qp->sq.max_gs;
391 
392 	/* We don't support inline sends for kernel QPs (yet) */
393 	cap->max_inline_data = 0;
394 
395 	return 0;
396 }
397 
398 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
399 				     struct ib_pd *ib_pd,
400 				     struct ib_qp_init_attr *init_attr,
401 				     struct ib_udata *udata, unsigned long sqpn,
402 				     struct hns_roce_qp *hr_qp)
403 {
404 	struct device *dev = &hr_dev->pdev->dev;
405 	struct hns_roce_ib_create_qp ucmd;
406 	unsigned long qpn = 0;
407 	int ret = 0;
408 
409 	mutex_init(&hr_qp->mutex);
410 	spin_lock_init(&hr_qp->sq.lock);
411 	spin_lock_init(&hr_qp->rq.lock);
412 
413 	hr_qp->state = IB_QPS_RESET;
414 
415 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
416 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
417 	else
418 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
419 
420 	ret = hns_roce_set_rq_size(hr_dev, &init_attr->cap, !!ib_pd->uobject,
421 				   !!init_attr->srq, hr_qp);
422 	if (ret) {
423 		dev_err(dev, "hns_roce_set_rq_size failed\n");
424 		goto err_out;
425 	}
426 
427 	if (ib_pd->uobject) {
428 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
429 			dev_err(dev, "ib_copy_from_udata error for create qp\n");
430 			ret = -EFAULT;
431 			goto err_out;
432 		}
433 
434 		ret = hns_roce_set_user_sq_size(hr_dev, hr_qp, &ucmd);
435 		if (ret) {
436 			dev_err(dev, "hns_roce_set_user_sq_size error for create qp\n");
437 			goto err_out;
438 		}
439 
440 		hr_qp->umem = ib_umem_get(ib_pd->uobject->context,
441 					  ucmd.buf_addr, hr_qp->buff_size, 0,
442 					  0);
443 		if (IS_ERR(hr_qp->umem)) {
444 			dev_err(dev, "ib_umem_get error for create qp\n");
445 			ret = PTR_ERR(hr_qp->umem);
446 			goto err_out;
447 		}
448 
449 		ret = hns_roce_mtt_init(hr_dev, ib_umem_page_count(hr_qp->umem),
450 				    ilog2((unsigned int)hr_qp->umem->page_size),
451 				    &hr_qp->mtt);
452 		if (ret) {
453 			dev_err(dev, "hns_roce_mtt_init error for create qp\n");
454 			goto err_buf;
455 		}
456 
457 		ret = hns_roce_ib_umem_write_mtt(hr_dev, &hr_qp->mtt,
458 						 hr_qp->umem);
459 		if (ret) {
460 			dev_err(dev, "hns_roce_ib_umem_write_mtt error for create qp\n");
461 			goto err_mtt;
462 		}
463 	} else {
464 		if (init_attr->create_flags &
465 		    IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
466 			dev_err(dev, "init_attr->create_flags error!\n");
467 			ret = -EINVAL;
468 			goto err_out;
469 		}
470 
471 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO) {
472 			dev_err(dev, "init_attr->create_flags error!\n");
473 			ret = -EINVAL;
474 			goto err_out;
475 		}
476 
477 		/* Set SQ size */
478 		ret = hns_roce_set_kernel_sq_size(hr_dev, &init_attr->cap,
479 						  init_attr->qp_type, hr_qp);
480 		if (ret) {
481 			dev_err(dev, "hns_roce_set_kernel_sq_size error!\n");
482 			goto err_out;
483 		}
484 
485 		/* QP doorbell register address */
486 		hr_qp->sq.db_reg_l = hr_dev->reg_base + ROCEE_DB_SQ_L_0_REG +
487 				     DB_REG_OFFSET * hr_dev->priv_uar.index;
488 		hr_qp->rq.db_reg_l = hr_dev->reg_base +
489 				     ROCEE_DB_OTHERS_L_0_REG +
490 				     DB_REG_OFFSET * hr_dev->priv_uar.index;
491 
492 		/* Allocate QP buf */
493 		if (hns_roce_buf_alloc(hr_dev, hr_qp->buff_size, PAGE_SIZE * 2,
494 				       &hr_qp->hr_buf)) {
495 			dev_err(dev, "hns_roce_buf_alloc error!\n");
496 			ret = -ENOMEM;
497 			goto err_out;
498 		}
499 
500 		/* Write MTT */
501 		ret = hns_roce_mtt_init(hr_dev, hr_qp->hr_buf.npages,
502 					hr_qp->hr_buf.page_shift, &hr_qp->mtt);
503 		if (ret) {
504 			dev_err(dev, "hns_roce_mtt_init error for kernel create qp\n");
505 			goto err_buf;
506 		}
507 
508 		ret = hns_roce_buf_write_mtt(hr_dev, &hr_qp->mtt,
509 					     &hr_qp->hr_buf);
510 		if (ret) {
511 			dev_err(dev, "hns_roce_buf_write_mtt error for kernel create qp\n");
512 			goto err_mtt;
513 		}
514 
515 		hr_qp->sq.wrid = kmalloc_array(hr_qp->sq.wqe_cnt, sizeof(u64),
516 					       GFP_KERNEL);
517 		hr_qp->rq.wrid = kmalloc_array(hr_qp->rq.wqe_cnt, sizeof(u64),
518 					       GFP_KERNEL);
519 		if (!hr_qp->sq.wrid || !hr_qp->rq.wrid) {
520 			ret = -ENOMEM;
521 			goto err_wrid;
522 		}
523 	}
524 
525 	if (sqpn) {
526 		qpn = sqpn;
527 	} else {
528 		/* Get QPN */
529 		ret = hns_roce_reserve_range_qp(hr_dev, 1, 1, &qpn);
530 		if (ret) {
531 			dev_err(dev, "hns_roce_reserve_range_qp alloc qpn error\n");
532 			goto err_wrid;
533 		}
534 	}
535 
536 	if ((init_attr->qp_type) == IB_QPT_GSI) {
537 		ret = hns_roce_gsi_qp_alloc(hr_dev, qpn, hr_qp);
538 		if (ret) {
539 			dev_err(dev, "hns_roce_qp_alloc failed!\n");
540 			goto err_qpn;
541 		}
542 	} else {
543 		ret = hns_roce_qp_alloc(hr_dev, qpn, hr_qp);
544 		if (ret) {
545 			dev_err(dev, "hns_roce_qp_alloc failed!\n");
546 			goto err_qpn;
547 		}
548 	}
549 
550 	if (sqpn)
551 		hr_qp->doorbell_qpn = 1;
552 	else
553 		hr_qp->doorbell_qpn = cpu_to_le64(hr_qp->qpn);
554 
555 	hr_qp->event = hns_roce_ib_qp_event;
556 
557 	return 0;
558 
559 err_qpn:
560 	if (!sqpn)
561 		hns_roce_release_range_qp(hr_dev, qpn, 1);
562 
563 err_wrid:
564 	kfree(hr_qp->sq.wrid);
565 	kfree(hr_qp->rq.wrid);
566 
567 err_mtt:
568 	hns_roce_mtt_cleanup(hr_dev, &hr_qp->mtt);
569 
570 err_buf:
571 	if (ib_pd->uobject)
572 		ib_umem_release(hr_qp->umem);
573 	else
574 		hns_roce_buf_free(hr_dev, hr_qp->buff_size, &hr_qp->hr_buf);
575 
576 err_out:
577 	return ret;
578 }
579 
580 struct ib_qp *hns_roce_create_qp(struct ib_pd *pd,
581 				 struct ib_qp_init_attr *init_attr,
582 				 struct ib_udata *udata)
583 {
584 	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
585 	struct device *dev = &hr_dev->pdev->dev;
586 	struct hns_roce_sqp *hr_sqp;
587 	struct hns_roce_qp *hr_qp;
588 	int ret;
589 
590 	switch (init_attr->qp_type) {
591 	case IB_QPT_RC: {
592 		hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
593 		if (!hr_qp)
594 			return ERR_PTR(-ENOMEM);
595 
596 		ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, 0,
597 						hr_qp);
598 		if (ret) {
599 			dev_err(dev, "Create RC QP failed\n");
600 			kfree(hr_qp);
601 			return ERR_PTR(ret);
602 		}
603 
604 		hr_qp->ibqp.qp_num = hr_qp->qpn;
605 
606 		break;
607 	}
608 	case IB_QPT_GSI: {
609 		/* Userspace is not allowed to create special QPs: */
610 		if (pd->uobject) {
611 			dev_err(dev, "not support usr space GSI\n");
612 			return ERR_PTR(-EINVAL);
613 		}
614 
615 		hr_sqp = kzalloc(sizeof(*hr_sqp), GFP_KERNEL);
616 		if (!hr_sqp)
617 			return ERR_PTR(-ENOMEM);
618 
619 		hr_qp = &hr_sqp->hr_qp;
620 
621 		ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata,
622 						hr_dev->caps.sqp_start +
623 						hr_dev->caps.num_ports +
624 						init_attr->port_num - 1, hr_qp);
625 		if (ret) {
626 			dev_err(dev, "Create GSI QP failed!\n");
627 			kfree(hr_sqp);
628 			return ERR_PTR(ret);
629 		}
630 
631 		hr_qp->port = (init_attr->port_num - 1);
632 		hr_qp->ibqp.qp_num = hr_dev->caps.sqp_start +
633 				     hr_dev->caps.num_ports +
634 				     init_attr->port_num - 1;
635 		break;
636 	}
637 	default:{
638 		dev_err(dev, "not support QP type %d\n", init_attr->qp_type);
639 		return ERR_PTR(-EINVAL);
640 	}
641 	}
642 
643 	return &hr_qp->ibqp;
644 }
645 
646 int to_hr_qp_type(int qp_type)
647 {
648 	int transport_type;
649 
650 	if (qp_type == IB_QPT_RC)
651 		transport_type = SERV_TYPE_RC;
652 	else if (qp_type == IB_QPT_UC)
653 		transport_type = SERV_TYPE_UC;
654 	else if (qp_type == IB_QPT_UD)
655 		transport_type = SERV_TYPE_UD;
656 	else if (qp_type == IB_QPT_GSI)
657 		transport_type = SERV_TYPE_UD;
658 	else
659 		transport_type = -1;
660 
661 	return transport_type;
662 }
663 
664 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
665 		       int attr_mask, struct ib_udata *udata)
666 {
667 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
668 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
669 	enum ib_qp_state cur_state, new_state;
670 	struct device *dev = &hr_dev->pdev->dev;
671 	int ret = -EINVAL;
672 	int p;
673 
674 	mutex_lock(&hr_qp->mutex);
675 
676 	cur_state = attr_mask & IB_QP_CUR_STATE ?
677 		    attr->cur_qp_state : (enum ib_qp_state)hr_qp->state;
678 	new_state = attr_mask & IB_QP_STATE ?
679 		    attr->qp_state : cur_state;
680 
681 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
682 				IB_LINK_LAYER_ETHERNET)) {
683 		dev_err(dev, "ib_modify_qp_is_ok failed\n");
684 		goto out;
685 	}
686 
687 	if ((attr_mask & IB_QP_PORT) &&
688 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
689 		dev_err(dev, "attr port_num invalid.attr->port_num=%d\n",
690 			attr->port_num);
691 		goto out;
692 	}
693 
694 	if (attr_mask & IB_QP_PKEY_INDEX) {
695 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
696 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
697 			dev_err(dev, "attr pkey_index invalid.attr->pkey_index=%d\n",
698 				attr->pkey_index);
699 			goto out;
700 		}
701 	}
702 
703 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
704 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
705 		dev_err(dev, "attr max_rd_atomic invalid.attr->max_rd_atomic=%d\n",
706 			attr->max_rd_atomic);
707 		goto out;
708 	}
709 
710 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
711 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
712 		dev_err(dev, "attr max_dest_rd_atomic invalid.attr->max_dest_rd_atomic=%d\n",
713 			attr->max_dest_rd_atomic);
714 		goto out;
715 	}
716 
717 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
718 		ret = -EPERM;
719 		dev_err(dev, "cur_state=%d new_state=%d\n", cur_state,
720 			new_state);
721 		goto out;
722 	}
723 
724 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
725 				    new_state);
726 
727 out:
728 	mutex_unlock(&hr_qp->mutex);
729 
730 	return ret;
731 }
732 
733 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
734 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
735 {
736 	if (send_cq == recv_cq) {
737 		spin_lock_irq(&send_cq->lock);
738 		__acquire(&recv_cq->lock);
739 	} else if (send_cq->cqn < recv_cq->cqn) {
740 		spin_lock_irq(&send_cq->lock);
741 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
742 	} else {
743 		spin_lock_irq(&recv_cq->lock);
744 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
745 	}
746 }
747 
748 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
749 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
750 			 __releases(&recv_cq->lock)
751 {
752 	if (send_cq == recv_cq) {
753 		__release(&recv_cq->lock);
754 		spin_unlock_irq(&send_cq->lock);
755 	} else if (send_cq->cqn < recv_cq->cqn) {
756 		spin_unlock(&recv_cq->lock);
757 		spin_unlock_irq(&send_cq->lock);
758 	} else {
759 		spin_unlock(&send_cq->lock);
760 		spin_unlock_irq(&recv_cq->lock);
761 	}
762 }
763 
764 __be32 send_ieth(struct ib_send_wr *wr)
765 {
766 	switch (wr->opcode) {
767 	case IB_WR_SEND_WITH_IMM:
768 	case IB_WR_RDMA_WRITE_WITH_IMM:
769 		return cpu_to_le32(wr->ex.imm_data);
770 	case IB_WR_SEND_WITH_INV:
771 		return cpu_to_le32(wr->ex.invalidate_rkey);
772 	default:
773 		return 0;
774 	}
775 }
776 
777 static void *get_wqe(struct hns_roce_qp *hr_qp, int offset)
778 {
779 
780 	return hns_roce_buf_offset(&hr_qp->hr_buf, offset);
781 }
782 
783 void *get_recv_wqe(struct hns_roce_qp *hr_qp, int n)
784 {
785 	struct ib_qp *ibqp = &hr_qp->ibqp;
786 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
787 
788 	if ((n < 0) || (n > hr_qp->rq.wqe_cnt)) {
789 		dev_err(&hr_dev->pdev->dev, "rq wqe index:%d,rq wqe cnt:%d\r\n",
790 			n, hr_qp->rq.wqe_cnt);
791 		return NULL;
792 	}
793 
794 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
795 }
796 
797 void *get_send_wqe(struct hns_roce_qp *hr_qp, int n)
798 {
799 	struct ib_qp *ibqp = &hr_qp->ibqp;
800 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
801 
802 	if ((n < 0) || (n > hr_qp->sq.wqe_cnt)) {
803 		dev_err(&hr_dev->pdev->dev, "sq wqe index:%d,sq wqe cnt:%d\r\n",
804 			n, hr_qp->sq.wqe_cnt);
805 		return NULL;
806 	}
807 
808 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
809 }
810 
811 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, int nreq,
812 			  struct ib_cq *ib_cq)
813 {
814 	struct hns_roce_cq *hr_cq;
815 	u32 cur;
816 
817 	cur = hr_wq->head - hr_wq->tail;
818 	if (likely(cur + nreq < hr_wq->max_post))
819 		return 0;
820 
821 	hr_cq = to_hr_cq(ib_cq);
822 	spin_lock(&hr_cq->lock);
823 	cur = hr_wq->head - hr_wq->tail;
824 	spin_unlock(&hr_cq->lock);
825 
826 	return cur + nreq >= hr_wq->max_post;
827 }
828 
829 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
830 {
831 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
832 	int reserved_from_top = 0;
833 	int ret;
834 
835 	spin_lock_init(&qp_table->lock);
836 	INIT_RADIX_TREE(&hr_dev->qp_table_tree, GFP_ATOMIC);
837 
838 	/* A port include two SQP, six port total 12 */
839 	ret = hns_roce_bitmap_init(&qp_table->bitmap, hr_dev->caps.num_qps,
840 				   hr_dev->caps.num_qps - 1,
841 				   hr_dev->caps.sqp_start + SQP_NUM,
842 				   reserved_from_top);
843 	if (ret) {
844 		dev_err(&hr_dev->pdev->dev, "qp bitmap init failed!error=%d\n",
845 			ret);
846 		return ret;
847 	}
848 
849 	return 0;
850 }
851 
852 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
853 {
854 	hns_roce_bitmap_cleanup(&hr_dev->qp_table.bitmap);
855 }
856